Ebtesam

AI/SE Research | CS PhD @GeorgeMasonU | Prev @MSFTResearch

[Jia-Bin Huang @jbhuang0604]

Scrolling the AI news timeline as a researcher feels like a teenager browsing Instagram:

"Everyone else has figured everything out!"

Reliable home robots imminent, 100× productivity AI agents, insane visual generation ...

Exciting, but anxiety-inducing. What am I doing? 😬

[Omar Khattab @lateinteraction]

crazy that they called it context window when attention span was right there

[Sebastian Raschka @rasbt]

As we all know by now, reasoning models often generate longer responses, which raises compute costs. Now, this new paper (https://t.co/UbBv4rzM09) shows that this behavior comes from the RL training process, not from an actual need for long answers for better accuracy. The RL loss tends to favor longer responses when the model gets negative rewards, which I think explains the "aha" moments and longer chains of thought that arise from pure RL training.

I.e., if the model gets a negative reward (i.e., the answer is wrong), the math behind PPO causes the average per-token loss becomes smaller when the response is longer. So, the model is indirectly encouraged to make its responses longer. This is true even if those extra tokens don't actually help solve the problem.

What does the response length have to do with the loss? When the reward is negative, longer responses can dilute the penalty per individual token, which results in lower (i.e., better) loss values (even though the model is still getting the answer wrong).

So the model "learns" that longer responses reduce the punishment, even though they are not helping correctness.

In addition, the researchers show that a second round of RL (using just a few problems that are sometimes solvable) can shorten responses while preserving or even improving accuracy. This has big implications for deployment efficiency.

[I Am Devloper @iamdevloper]

vibe coding, where 2 engineers can now create the tech debt of at least 50 engineers

[Nabeel S. Qureshi @nabeelqu]

For the confused, it's actually super easy:
- GPT 4.5 is the new Claude 3.6 (aka 3.5)
- Claude 3.7 is the new o3-mini-high
- Claude Code is the new Cursor
- Grok is the new Perplexity
- o1 pro is the 'smartest', except for o3, which backs Deep Research

Obviously. Keep up.

[Hamel Husain @HamelHusain]

New post re: Devin (the AI SWE). We couldn't find many reviews of people using it for real tasks, so we went MKBHD mode and put Devin through its paces.

We documented our findings here. Would love to know if others have had a different experience.

https://t.co/DDqzoAXKkl

[Diomidis Spinellis @CoolSWEng]

Long overdue, a paper finally exposes the Emperor's New “Threats to Validity” Clothes in empirical software engineering research. Even better, it provides suggestions for improving the state of practice.

[Jiaxin Pei @jiaxin_pei]

It's common to add personas in system prompts, assuming this can help LLMs. However, through analyzing 162 roles x 4 LLMs x 2410 questions, we show that adding a persona mostly has *no* statistically significant difference from the no-persona setting. If there is a difference, it is *negative*. It's time to rethink the usage of personas in system prompts!

[Ishan @radshaan]

If you get frequent urges to go deep into a subject, do not ignore them

Pick a weekend, stop everything else, and give in to the urge

Fresh insights await at the other end

[Upol Ehsan @UpolEhsan]

Is hallucination in LLMs inevitable even with an idealized model architecture and perfect training data?

This work argues YES and offers a formal proof.

Let's dig in ⤵

🧵1/n

[Edward Grefenstette @egrefen]

Instead, evaluation processes should track the diverse notions of extrinsic utility which are to be found in both everyday usage of our technology today, but also anticipating how people might use technology tomorrow.

[Dr Meming @Dr_Meming]

Heck

[INSPIRED Lab @ GMU @INSPIREDLabGMU]

🚨 Inclusive tech research alert! 🚨

Are you a tech user who identifies as BIPOC (https://t.co/gj9uuPIz4d)? Or a researcher/practitioner who uses data in your work?

Share your experiences in our 20 min. survey→​https://t.co/rGBpQUChFO

IRBNet #: 1945546-2
#data #tech #trust

[Dr. Amy Lee @minisciencegirl]

Never name a manuscript draft "_FINAL"

[Dr Meming @Dr_Meming]

Academic research: months of experiments and data analysis that ends up being a few sentences in a paper

[will depue @willdepue]

I feel like large language model feels a bit reductive when GPT-2 is in the same class as GPT-4. gigantic language models? enormous language models? big ass language models? Nimitz-class language models? better suggestions needed

[MIT CSAIL @MIT_CSAIL]

Happy birthday to Python creator Guido van Rossum. The open source language was named after comedy troupe Monty Python: https://t.co/UGUO3rp0M1

Image v/Midjourney

[François Chollet @fchollet]

When I got started with programming, I debugged using printf() statements. Today, I debug with print() statements.

The purpose of debugging is to correct your mental model of what your code does, and no tool can do that for you. The best any tool can do is provide visibility into code execution, and targeted print statements already do a tremendous job at that.